Departement Mathematik und Informatik / Computergraphik Bilderkennung (Vetter)
We use the 3DMM to analyze images of human faces in an Analysis-by-Synthesis manner. The model is used to actively reconstruct the 3D shape of the depicted faces from a single image. This analysis task is one of the main applications of the Morphable Model of faces since the beginning. But so far, the analysis needed manual user intervention to initialize the model appropriately. An automatic initialization is possible using face and feature point detection technology but it has always been unreliable for non-frontal faces. In this project focus, we develop detection methods which can deal with non-frontal situations. To deal with unreliable input information, we directly integrate detection information into our probabilistic Markov Chain Monte Carlo model fitting method. The method is based on a propose-and-verify architecture which can deal with unreliable information by separating proposals from their verification with the generative model.